block.c

   1 /*
   2  * QEMU System Emulator block driver
   3  *
   4  * Copyright (c) 2003 Fabrice Bellard
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24 #include "config-host.h"
  25 #ifdef HOST_BSD
  26 /* include native header before sys-queue.h */
  27 #include <sys/queue.h>
  28 #endif
  29
  30 #include "qemu-common.h"
  31 #include "monitor.h"
  32 #include "block_int.h"
  33 #include "module.h"
  34
  35 #ifdef HOST_BSD
  36 #include <sys/types.h>
  37 #include <sys/stat.h>
  38 #include <sys/ioctl.h>
  39 #ifndef __DragonFly__
  40 #include <sys/disk.h>
  41 #endif
  42 #endif
  43
  44 #ifdef _WIN32
  45 #include <windows.h>
  46 #endif
  47
  48 #define SECTOR_BITS 9
  49 #define SECTOR_SIZE (1 << SECTOR_BITS)
  50
  51 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
  52         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  53         BlockDriverCompletionFunc *cb, void *opaque);
  54 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
  55         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  56         BlockDriverCompletionFunc *cb, void *opaque);
  57 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  58                         uint8_t *buf, int nb_sectors);
  59 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  60                          const uint8_t *buf, int nb_sectors);
  61
  62 BlockDriverState *bdrv_first;
  63
  64 static BlockDriver *first_drv;
  65
  66 int path_is_absolute(const char *path)
  67 {
  68     const char *p;
  69 #ifdef _WIN32
  70     /* specific case for names like: "\\.\d:" */
  71     if (*path == '/' || *path == '\\')
  72         return 1;
  73 #endif
  74     p = strchr(path, ':');
  75     if (p)
  76         p++;
  77     else
  78         p = path;
  79 #ifdef _WIN32
  80     return (*p == '/' || *p == '\\');
  81 #else
  82     return (*p == '/');
  83 #endif
  84 }
  85
  86 /* if filename is absolute, just copy it to dest. Otherwise, build a
  87    path to it by considering it is relative to base_path. URL are
  88    supported. */
  89 void path_combine(char *dest, int dest_size,
  90                   const char *base_path,
  91                   const char *filename)
  92 {
  93     const char *p, *p1;
  94     int len;
  95
  96     if (dest_size <= 0)
  97         return;
  98     if (path_is_absolute(filename)) {
  99         pstrcpy(dest, dest_size, filename);
 100     } else {
 101         p = strchr(base_path, ':');
 102         if (p)
 103             p++;
 104         else
 105             p = base_path;
 106         p1 = strrchr(base_path, '/');
 107 #ifdef _WIN32
 108         {
 109             const char *p2;
 110             p2 = strrchr(base_path, '\\');
 111             if (!p1 || p2 > p1)
 112                 p1 = p2;
 113         }
 114 #endif
 115         if (p1)
 116             p1++;
 117         else
 118             p1 = base_path;
 119         if (p1 > p)
 120             p = p1;
 121         len = p - base_path;
 122         if (len > dest_size - 1)
 123             len = dest_size - 1;
 124         memcpy(dest, base_path, len);
 125         dest[len] = '\0';
 126         pstrcat(dest, dest_size, filename);
 127     }
 128 }
 129
 130 void bdrv_register(BlockDriver *bdrv)
 131 {
 132     if (!bdrv->bdrv_aio_readv) {
 133         /* add AIO emulation layer */
 134         bdrv->bdrv_aio_readv = bdrv_aio_readv_em;
 135         bdrv->bdrv_aio_writev = bdrv_aio_writev_em;
 136     } else if (!bdrv->bdrv_read) {
 137         /* add synchronous IO emulation layer */
 138         bdrv->bdrv_read = bdrv_read_em;
 139         bdrv->bdrv_write = bdrv_write_em;
 140     }
 141     bdrv->next = first_drv;
 142     first_drv = bdrv;
 143 }
 144
 145 /* create a new block device (by default it is empty) */
 146 BlockDriverState *bdrv_new(const char *device_name)
 147 {
 148     BlockDriverState **pbs, *bs;
 149
 150     bs = qemu_mallocz(sizeof(BlockDriverState));
 151     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 152     if (device_name[0] != '\0') {
 153         /* insert at the end */
 154         pbs = &bdrv_first;
 155         while (*pbs != NULL)
 156             pbs = &(*pbs)->next;
 157         *pbs = bs;
 158     }
 159     return bs;
 160 }
 161
 162 BlockDriver *bdrv_find_format(const char *format_name)
 163 {
 164     BlockDriver *drv1;
 165     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 166         if (!strcmp(drv1->format_name, format_name))
 167             return drv1;
 168     }
 169     return NULL;
 170 }
 171
 172 int bdrv_create(BlockDriver *drv, const char* filename,
 173     QEMUOptionParameter *options)
 174 {
 175     if (!drv->bdrv_create)
 176         return -ENOTSUP;
 177
 178     return drv->bdrv_create(filename, options);
 179 }
 180
 181 #ifdef _WIN32
 182 void get_tmp_filename(char *filename, int size)
 183 {
 184     char temp_dir[MAX_PATH];
 185
 186     GetTempPath(MAX_PATH, temp_dir);
 187     GetTempFileName(temp_dir, "qem", 0, filename);
 188 }
 189 #else
 190 void get_tmp_filename(char *filename, int size)
 191 {
 192     int fd;
 193     const char *tmpdir;
 194     /* XXX: race condition possible */
 195     tmpdir = getenv("TMPDIR");
 196     if (!tmpdir)
 197         tmpdir = "/tmp";
 198     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 199     fd = mkstemp(filename);
 200     close(fd);
 201 }
 202 #endif
 203
 204 #ifdef _WIN32
 205 static int is_windows_drive_prefix(const char *filename)
 206 {
 207     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 208              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 209             filename[1] == ':');
 210 }
 211
 212 int is_windows_drive(const char *filename)
 213 {
 214     if (is_windows_drive_prefix(filename) &&
 215         filename[2] == '\0')
 216         return 1;
 217     if (strstart(filename, "\\\\.\\", NULL) ||
 218         strstart(filename, "//./", NULL))
 219         return 1;
 220     return 0;
 221 }
 222 #endif
 223
 224 static BlockDriver *find_protocol(const char *filename)
 225 {
 226     BlockDriver *drv1;
 227     char protocol[128];
 228     int len = strnlen(filename, 127)+1;
 229     const char *p;
 230
 231 #ifdef _WIN32
 232     if (is_windows_drive(filename) ||
 233         is_windows_drive_prefix(filename))
 234         return bdrv_find_format("raw");
 235 #endif
 236     p = fill_token(protocol, len, filename, ':');
 237     if (*p != ':')
 238         return bdrv_find_format("raw");
 239     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 240         if (drv1->protocol_name &&
 241             !strcmp(drv1->protocol_name, protocol))
 242             return drv1;
 243     }
 244     return NULL;
 245 }
 246
 247 /*
 248  * Detect host devices. By convention, /dev/cdrom[N] is always
 249  * recognized as a host CDROM.
 250  */
 251 static BlockDriver *find_hdev_driver(const char *filename)
 252 {
 253     int score_max = 0, score;
 254     BlockDriver *drv = NULL, *d;
 255
 256     for (d = first_drv; d; d = d->next) {
 257         if (d->bdrv_probe_device) {
 258             score = d->bdrv_probe_device(filename);
 259             if (score > score_max) {
 260                 score_max = score;
 261                 drv = d;
 262             }
 263         }
 264     }
 265
 266     return drv;
 267 }
 268
 269 static BlockDriver *find_image_format(const char *filename)
 270 {
 271     int ret, score, score_max;
 272     BlockDriver *drv1, *drv;
 273     uint8_t buf[2048];
 274     BlockDriverState *bs;
 275
 276     drv = find_protocol(filename);
 277     /* no need to test disk image formats for vvfat */
 278     if (drv && strcmp(drv->format_name, "vvfat") == 0)
 279         return drv;
 280
 281     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 282     if (ret < 0)
 283         return NULL;
 284     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 285     bdrv_delete(bs);
 286     if (ret < 0) {
 287         return NULL;
 288     }
 289
 290     score_max = 0;
 291     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 292         if (drv1->bdrv_probe) {
 293             score = drv1->bdrv_probe(buf, ret, filename);
 294             if (score > score_max) {
 295                 score_max = score;
 296                 drv = drv1;
 297             }
 298         }
 299     }
 300     return drv;
 301 }
 302
 303 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 304 {
 305     BlockDriverState *bs;
 306     int ret;
 307
 308     bs = bdrv_new("");
 309     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 310     if (ret < 0) {
 311         bdrv_delete(bs);
 312         return ret;
 313     }
 314     bs->growable = 1;
 315     *pbs = bs;
 316     return 0;
 317 }
 318
 319 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 320 {
 321     return bdrv_open2(bs, filename, flags, NULL);
 322 }
 323
 324 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 325                BlockDriver *drv)
 326 {
 327     int ret, open_flags;
 328     char tmp_filename[PATH_MAX];
 329     char backing_filename[PATH_MAX];
 330
 331     bs->read_only = 0;
 332     bs->is_temporary = 0;
 333     bs->encrypted = 0;
 334     bs->valid_key = 0;
 335     /* buffer_alignment defaulted to 512, drivers can change this value */
 336     bs->buffer_alignment = 512;
 337
 338     if (flags & BDRV_O_SNAPSHOT) {
 339         BlockDriverState *bs1;
 340         int64_t total_size;
 341         int is_protocol = 0;
 342         BlockDriver *bdrv_qcow2;
 343         QEMUOptionParameter *options;
 344
 345         /* if snapshot, we create a temporary backing file and open it
 346            instead of opening 'filename' directly */
 347
 348         /* if there is a backing file, use it */
 349         bs1 = bdrv_new("");
 350         ret = bdrv_open2(bs1, filename, 0, drv);
 351         if (ret < 0) {
 352             bdrv_delete(bs1);
 353             return ret;
 354         }
 355         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 356
 357         if (bs1->drv && bs1->drv->protocol_name)
 358             is_protocol = 1;
 359
 360         bdrv_delete(bs1);
 361
 362         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 363
 364         /* Real path is meaningless for protocols */
 365         if (is_protocol)
 366             snprintf(backing_filename, sizeof(backing_filename),
 367                      "%s", filename);
 368         else
 369             realpath(filename, backing_filename);
 370
 371         bdrv_qcow2 = bdrv_find_format("qcow2");
 372         options = parse_option_parameters("", bdrv_qcow2->create_options, NULL);
 373
 374         set_option_parameter_int(options, BLOCK_OPT_SIZE, total_size * 512);
 375         set_option_parameter(options, BLOCK_OPT_BACKING_FILE, backing_filename);
 376         if (drv) {
 377             set_option_parameter(options, BLOCK_OPT_BACKING_FMT,
 378                 drv->format_name);
 379         }
 380
 381         ret = bdrv_create(bdrv_qcow2, tmp_filename, options);
 382         if (ret < 0) {
 383             return ret;
 384         }
 385
 386         filename = tmp_filename;
 387         drv = bdrv_qcow2;
 388         bs->is_temporary = 1;
 389     }
 390
 391     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 392     if (flags & BDRV_O_FILE) {
 393         drv = find_protocol(filename);
 394     } else if (!drv) {
 395         drv = find_hdev_driver(filename);
 396         if (!drv) {
 397             drv = find_image_format(filename);
 398         }
 399     }
 400     if (!drv) {
 401         ret = -ENOENT;
 402         goto unlink_and_fail;
 403     }
 404     bs->drv = drv;
 405     bs->opaque = qemu_mallocz(drv->instance_size);
 406     /* Note: for compatibility, we open disk image files as RDWR, and
 407        RDONLY as fallback */
 408     if (!(flags & BDRV_O_FILE))
 409         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 410     else
 411         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 412     ret = bdrv_open3(bs, filename, open_flags, drv);
 413     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 414         ret = bdrv_open3(bs, filename, open_flags & ~BDRV_O_RDWR, drv);
 415         bs->read_only = 1;
 416     }
 417     if (ret < 0) {
 418         qemu_free(bs->opaque);
 419         bs->opaque = NULL;
 420         bs->drv = NULL;
 421     unlink_and_fail:
 422         if (bs->is_temporary)
 423             unlink(filename);
 424         return ret;
 425     }
 426     if (drv->bdrv_getlength) {
 427         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 428     }
 429 #ifndef _WIN32
 430     if (bs->is_temporary) {
 431         unlink(filename);
 432     }
 433 #endif
 434     if (bs->backing_file[0] != '\0') {
 435         /* if there is a backing file, use it */
 436         BlockDriver *back_drv = NULL;
 437         bs->backing_hd = bdrv_new("");
 438         path_combine(backing_filename, sizeof(backing_filename),
 439                      filename, bs->backing_file);
 440         if (bs->backing_format[0] != '\0')
 441             back_drv = bdrv_find_format(bs->backing_format);
 442         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 443                          back_drv);
 444         if (ret < 0) {
 445             bdrv_close(bs);
 446             return ret;
 447         }
 448     }
 449
 450     if (!bdrv_key_required(bs)) {
 451         /* call the change callback */
 452         bs->media_changed = 1;
 453         if (bs->change_cb)
 454             bs->change_cb(bs->change_opaque);
 455     }
 456     return 0;
 457 }
 458
 459 int bdrv_open3(BlockDriverState *bs, const char *filename, int flags, BlockDriver *drv)
 460 {
 461     char myfile[PATH_MAX];
 462     const char *f;
 463
 464     if (!strstart(filename, "file:", &f)) {
 465         fill_token(myfile, PATH_MAX, filename, '\0');
 466         return drv->bdrv_open(bs,myfile,flags);
 467     }
 468     return drv->bdrv_open(bs,f,flags);
 469 }
 470
 471 void bdrv_close(BlockDriverState *bs)
 472 {
 473     if (bs->drv) {
 474         if (bs->backing_hd)
 475             bdrv_delete(bs->backing_hd);
 476         bs->drv->bdrv_close(bs);
 477         qemu_free(bs->opaque);
 478 #ifdef _WIN32
 479         if (bs->is_temporary) {
 480             unlink(bs->filename);
 481         }
 482 #endif
 483         bs->opaque = NULL;
 484         bs->drv = NULL;
 485
 486         /* call the change callback */
 487         bs->media_changed = 1;
 488         if (bs->change_cb)
 489             bs->change_cb(bs->change_opaque);
 490     }
 491 }
 492
 493 void bdrv_delete(BlockDriverState *bs)
 494 {
 495     BlockDriverState **pbs;
 496
 497     pbs = &bdrv_first;
 498     while (*pbs != bs && *pbs != NULL)
 499         pbs = &(*pbs)->next;
 500     if (*pbs == bs)
 501         *pbs = bs->next;
 502
 503     bdrv_close(bs);
 504     qemu_free(bs);
 505 }
 506
 507 /*
 508  * Run consistency checks on an image
 509  *
 510  * Returns the number of errors or -errno when an internal error occurs
 511  */
 512 int bdrv_check(BlockDriverState *bs)
 513 {
 514     if (bs->drv->bdrv_check == NULL) {
 515         return -ENOTSUP;
 516     }
 517
 518     return bs->drv->bdrv_check(bs);
 519 }
 520
 521 /* commit COW file into the raw image */
 522 int bdrv_commit(BlockDriverState *bs)
 523 {
 524     BlockDriver *drv = bs->drv;
 525     int64_t i, total_sectors;
 526     int n, j;
 527     unsigned char sector[512];
 528
 529     if (!drv)
 530         return -ENOMEDIUM;
 531
 532     if (bs->read_only) {
 533         return -EACCES;
 534     }
 535
 536     if (!bs->backing_hd) {
 537         return -ENOTSUP;
 538     }
 539
 540     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 541     for (i = 0; i < total_sectors;) {
 542         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 543             for(j = 0; j < n; j++) {
 544                 if (bdrv_read(bs, i, sector, 1) != 0) {
 545                     return -EIO;
 546                 }
 547
 548                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 549                     return -EIO;
 550                 }
 551                 i++;
 552             }
 553         } else {
 554             i += n;
 555         }
 556     }
 557
 558     if (drv->bdrv_make_empty)
 559         return drv->bdrv_make_empty(bs);
 560
 561     return 0;
 562 }
 563
 564 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 565                                    size_t size)
 566 {
 567     int64_t len;
 568
 569     if (!bdrv_is_inserted(bs))
 570         return -ENOMEDIUM;
 571
 572     if (bs->growable)
 573         return 0;
 574
 575     len = bdrv_getlength(bs);
 576
 577     if (offset < 0)
 578         return -EIO;
 579
 580     if ((offset > len) || (len - offset < size))
 581         return -EIO;
 582
 583     return 0;
 584 }
 585
 586 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 587                               int nb_sectors)
 588 {
 589     return bdrv_check_byte_request(bs, sector_num * 512, nb_sectors * 512);
 590 }
 591
 592 /* return < 0 if error. See bdrv_write() for the return codes */
 593 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 594               uint8_t *buf, int nb_sectors)
 595 {
 596     BlockDriver *drv = bs->drv;
 597
 598     if (!drv)
 599         return -ENOMEDIUM;
 600     if (bdrv_check_request(bs, sector_num, nb_sectors))
 601         return -EIO;
 602
 603     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 604 }
 605
 606 /* Return < 0 if error. Important errors are:
 607   -EIO         generic I/O error (may happen for all errors)
 608   -ENOMEDIUM   No media inserted.
 609   -EINVAL      Invalid sector number or nb_sectors
 610   -EACCES      Trying to write a read-only device
 611 */
 612 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 613                const uint8_t *buf, int nb_sectors)
 614 {
 615     BlockDriver *drv = bs->drv;
 616     if (!bs->drv)
 617         return -ENOMEDIUM;
 618     if (bs->read_only)
 619         return -EACCES;
 620     if (bdrv_check_request(bs, sector_num, nb_sectors))
 621         return -EIO;
 622
 623     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 624 }
 625
 626 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 627                void *buf, int count1)
 628 {
 629     uint8_t tmp_buf[SECTOR_SIZE];
 630     int len, nb_sectors, count;
 631     int64_t sector_num;
 632
 633     count = count1;
 634     /* first read to align to sector start */
 635     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 636     if (len > count)
 637         len = count;
 638     sector_num = offset >> SECTOR_BITS;
 639     if (len > 0) {
 640         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 641             return -EIO;
 642         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 643         count -= len;
 644         if (count == 0)
 645             return count1;
 646         sector_num++;
 647         buf += len;
 648     }
 649
 650     /* read the sectors "in place" */
 651     nb_sectors = count >> SECTOR_BITS;
 652     if (nb_sectors > 0) {
 653         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 654             return -EIO;
 655         sector_num += nb_sectors;
 656         len = nb_sectors << SECTOR_BITS;
 657         buf += len;
 658         count -= len;
 659     }
 660
 661     /* add data from the last sector */
 662     if (count > 0) {
 663         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 664             return -EIO;
 665         memcpy(buf, tmp_buf, count);
 666     }
 667     return count1;
 668 }
 669
 670 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 671                 const void *buf, int count1)
 672 {
 673     uint8_t tmp_buf[SECTOR_SIZE];
 674     int len, nb_sectors, count;
 675     int64_t sector_num;
 676
 677     count = count1;
 678     /* first write to align to sector start */
 679     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 680     if (len > count)
 681         len = count;
 682     sector_num = offset >> SECTOR_BITS;
 683     if (len > 0) {
 684         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 685             return -EIO;
 686         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 687         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 688             return -EIO;
 689         count -= len;
 690         if (count == 0)
 691             return count1;
 692         sector_num++;
 693         buf += len;
 694     }
 695
 696     /* write the sectors "in place" */
 697     nb_sectors = count >> SECTOR_BITS;
 698     if (nb_sectors > 0) {
 699         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 700             return -EIO;
 701         sector_num += nb_sectors;
 702         len = nb_sectors << SECTOR_BITS;
 703         buf += len;
 704         count -= len;
 705     }
 706
 707     /* add data from the last sector */
 708     if (count > 0) {
 709         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 710             return -EIO;
 711         memcpy(tmp_buf, buf, count);
 712         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 713             return -EIO;
 714     }
 715     return count1;
 716 }
 717
 718 /**
 719  * Truncate file to 'offset' bytes (needed only for file protocols)
 720  */
 721 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 722 {
 723     BlockDriver *drv = bs->drv;
 724     if (!drv)
 725         return -ENOMEDIUM;
 726     if (!drv->bdrv_truncate)
 727         return -ENOTSUP;
 728     return drv->bdrv_truncate(bs, offset);
 729 }
 730
 731 /**
 732  * Length of a file in bytes. Return < 0 if error or unknown.
 733  */
 734 int64_t bdrv_getlength(BlockDriverState *bs)
 735 {
 736     BlockDriver *drv = bs->drv;
 737     if (!drv)
 738         return -ENOMEDIUM;
 739     if (!drv->bdrv_getlength) {
 740         /* legacy mode */
 741         return bs->total_sectors * SECTOR_SIZE;
 742     }
 743     return drv->bdrv_getlength(bs);
 744 }
 745
 746 /* return 0 as number of sectors if no device present or error */
 747 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 748 {
 749     int64_t length;
 750     length = bdrv_getlength(bs);
 751     if (length < 0)
 752         length = 0;
 753     else
 754         length = length >> SECTOR_BITS;
 755     *nb_sectors_ptr = length;
 756 }
 757
 758 struct partition {
 759         uint8_t boot_ind;           /* 0x80 - active */
 760         uint8_t head;               /* starting head */
 761         uint8_t sector;             /* starting sector */
 762         uint8_t cyl;                /* starting cylinder */
 763         uint8_t sys_ind;            /* What partition type */
 764         uint8_t end_head;           /* end head */
 765         uint8_t end_sector;         /* end sector */
 766         uint8_t end_cyl;            /* end cylinder */
 767         uint32_t start_sect;        /* starting sector counting from 0 */
 768         uint32_t nr_sects;          /* nr of sectors in partition */
 769 } __attribute__((packed));
 770
 771 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 772 static int guess_disk_lchs(BlockDriverState *bs,
 773                            int *pcylinders, int *pheads, int *psectors)
 774 {
 775     uint8_t buf[512];
 776     int ret, i, heads, sectors, cylinders;
 777     struct partition *p;
 778     uint32_t nr_sects;
 779     uint64_t nb_sectors;
 780
 781     bdrv_get_geometry(bs, &nb_sectors);
 782
 783     ret = bdrv_read(bs, 0, buf, 1);
 784     if (ret < 0)
 785         return -1;
 786     /* test msdos magic */
 787     if (buf[510] != 0x55 || buf[511] != 0xaa)
 788         return -1;
 789     for(i = 0; i < 4; i++) {
 790         p = ((struct partition *)(buf + 0x1be)) + i;
 791         nr_sects = le32_to_cpu(p->nr_sects);
 792         if (nr_sects && p->end_head) {
 793             /* We make the assumption that the partition terminates on
 794                a cylinder boundary */
 795             heads = p->end_head + 1;
 796             sectors = p->end_sector & 63;
 797             if (sectors == 0)
 798                 continue;
 799             cylinders = nb_sectors / (heads * sectors);
 800             if (cylinders < 1 || cylinders > 16383)
 801                 continue;
 802             *pheads = heads;
 803             *psectors = sectors;
 804             *pcylinders = cylinders;
 805 #if 0
 806             printf("guessed geometry: LCHS=%d %d %d\n",
 807                    cylinders, heads, sectors);
 808 #endif
 809             return 0;
 810         }
 811     }
 812     return -1;
 813 }
 814
 815 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 816 {
 817     int translation, lba_detected = 0;
 818     int cylinders, heads, secs;
 819     uint64_t nb_sectors;
 820
 821     /* if a geometry hint is available, use it */
 822     bdrv_get_geometry(bs, &nb_sectors);
 823     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 824     translation = bdrv_get_translation_hint(bs);
 825     if (cylinders != 0) {
 826         *pcyls = cylinders;
 827         *pheads = heads;
 828         *psecs = secs;
 829     } else {
 830         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 831             if (heads > 16) {
 832                 /* if heads > 16, it means that a BIOS LBA
 833                    translation was active, so the default
 834                    hardware geometry is OK */
 835                 lba_detected = 1;
 836                 goto default_geometry;
 837             } else {
 838                 *pcyls = cylinders;
 839                 *pheads = heads;
 840                 *psecs = secs;
 841                 /* disable any translation to be in sync with
 842                    the logical geometry */
 843                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 844                     bdrv_set_translation_hint(bs,
 845                                               BIOS_ATA_TRANSLATION_NONE);
 846                 }
 847             }
 848         } else {
 849         default_geometry:
 850             /* if no geometry, use a standard physical disk geometry */
 851             cylinders = nb_sectors / (16 * 63);
 852
 853             if (cylinders > 16383)
 854                 cylinders = 16383;
 855             else if (cylinders < 2)
 856                 cylinders = 2;
 857             *pcyls = cylinders;
 858             *pheads = 16;
 859             *psecs = 63;
 860             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 861                 if ((*pcyls * *pheads) <= 131072) {
 862                     bdrv_set_translation_hint(bs,
 863                                               BIOS_ATA_TRANSLATION_LARGE);
 864                 } else {
 865                     bdrv_set_translation_hint(bs,
 866                                               BIOS_ATA_TRANSLATION_LBA);
 867                 }
 868             }
 869         }
 870         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 871     }
 872 }
 873
 874 void bdrv_set_geometry_hint(BlockDriverState *bs,
 875                             int cyls, int heads, int secs)
 876 {
 877     bs->cyls = cyls;
 878     bs->heads = heads;
 879     bs->secs = secs;
 880 }
 881
 882 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 883 {
 884     bs->type = type;
 885     bs->removable = ((type == BDRV_TYPE_CDROM ||
 886                       type == BDRV_TYPE_FLOPPY));
 887 }
 888
 889 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 890 {
 891     bs->translation = translation;
 892 }
 893
 894 void bdrv_get_geometry_hint(BlockDriverState *bs,
 895                             int *pcyls, int *pheads, int *psecs)
 896 {
 897     *pcyls = bs->cyls;
 898     *pheads = bs->heads;
 899     *psecs = bs->secs;
 900 }
 901
 902 int bdrv_get_type_hint(BlockDriverState *bs)
 903 {
 904     return bs->type;
 905 }
 906
 907 int bdrv_get_translation_hint(BlockDriverState *bs)
 908 {
 909     return bs->translation;
 910 }
 911
 912 int bdrv_is_removable(BlockDriverState *bs)
 913 {
 914     return bs->removable;
 915 }
 916
 917 int bdrv_is_read_only(BlockDriverState *bs)
 918 {
 919     return bs->read_only;
 920 }
 921
 922 int bdrv_is_sg(BlockDriverState *bs)
 923 {
 924     return bs->sg;
 925 }
 926
 927 /* XXX: no longer used */
 928 void bdrv_set_change_cb(BlockDriverState *bs,
 929                         void (*change_cb)(void *opaque), void *opaque)
 930 {
 931     bs->change_cb = change_cb;
 932     bs->change_opaque = opaque;
 933 }
 934
 935 int bdrv_is_encrypted(BlockDriverState *bs)
 936 {
 937     if (bs->backing_hd && bs->backing_hd->encrypted)
 938         return 1;
 939     return bs->encrypted;
 940 }
 941
 942 int bdrv_key_required(BlockDriverState *bs)
 943 {
 944     BlockDriverState *backing_hd = bs->backing_hd;
 945
 946     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 947         return 1;
 948     return (bs->encrypted && !bs->valid_key);
 949 }
 950
 951 int bdrv_set_key(BlockDriverState *bs, const char *key)
 952 {
 953     int ret;
 954     if (bs->backing_hd && bs->backing_hd->encrypted) {
 955         ret = bdrv_set_key(bs->backing_hd, key);
 956         if (ret < 0)
 957             return ret;
 958         if (!bs->encrypted)
 959             return 0;
 960     }
 961     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 962         return -1;
 963     ret = bs->drv->bdrv_set_key(bs, key);
 964     if (ret < 0) {
 965         bs->valid_key = 0;
 966     } else if (!bs->valid_key) {
 967         bs->valid_key = 1;
 968         /* call the change callback now, we skipped it on open */
 969         bs->media_changed = 1;
 970         if (bs->change_cb)
 971             bs->change_cb(bs->change_opaque);
 972     }
 973     return ret;
 974 }
 975
 976 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 977 {
 978     if (!bs->drv) {
 979         buf[0] = '\0';
 980     } else {
 981         pstrcpy(buf, buf_size, bs->drv->format_name);
 982     }
 983 }
 984
 985 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 986                          void *opaque)
 987 {
 988     BlockDriver *drv;
 989
 990     for (drv = first_drv; drv != NULL; drv = drv->next) {
 991         it(opaque, drv->format_name);
 992     }
 993 }
 994
 995 BlockDriverState *bdrv_find(const char *name)
 996 {
 997     BlockDriverState *bs;
 998
 999     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1000         if (!strcmp(name, bs->device_name))
1001             return bs;
1002     }
1003     return NULL;
1004 }
1005
1006 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
1007 {
1008     BlockDriverState *bs;
1009
1010     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1011         it(opaque, bs);
1012     }
1013 }
1014
1015 const char *bdrv_get_device_name(BlockDriverState *bs)
1016 {
1017     return bs->device_name;
1018 }
1019
1020 void bdrv_flush(BlockDriverState *bs)
1021 {
1022     if (!bs->drv)
1023         return;
1024     if (bs->drv->bdrv_flush)
1025         bs->drv->bdrv_flush(bs);
1026     if (bs->backing_hd)
1027         bdrv_flush(bs->backing_hd);
1028 }
1029
1030 void bdrv_flush_all(void)
1031 {
1032     BlockDriverState *bs;
1033
1034     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1035         if (bs->drv && !bdrv_is_read_only(bs) &&
1036             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1037             bdrv_flush(bs);
1038 }
1039
1040 /*
1041  * Returns true iff the specified sector is present in the disk image. Drivers
1042  * not implementing the functionality are assumed to not support backing files,
1043  * hence all their sectors are reported as allocated.
1044  *
1045  * 'pnum' is set to the number of sectors (including and immediately following
1046  * the specified sector) that are known to be in the same
1047  * allocated/unallocated state.
1048  *
1049  * 'nb_sectors' is the max value 'pnum' should be set to.
1050  */
1051 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1052         int *pnum)
1053 {
1054     int64_t n;
1055     if (!bs->drv->bdrv_is_allocated) {
1056         if (sector_num >= bs->total_sectors) {
1057             *pnum = 0;
1058             return 0;
1059         }
1060         n = bs->total_sectors - sector_num;
1061         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1062         return 1;
1063     }
1064     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1065 }
1066
1067 void bdrv_info(Monitor *mon)
1068 {
1069     BlockDriverState *bs;
1070
1071     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1072         monitor_printf(mon, "%s:", bs->device_name);
1073         monitor_printf(mon, " type=");
1074         switch(bs->type) {
1075         case BDRV_TYPE_HD:
1076             monitor_printf(mon, "hd");
1077             break;
1078         case BDRV_TYPE_CDROM:
1079             monitor_printf(mon, "cdrom");
1080             break;
1081         case BDRV_TYPE_FLOPPY:
1082             monitor_printf(mon, "floppy");
1083             break;
1084         }
1085         monitor_printf(mon, " removable=%d", bs->removable);
1086         if (bs->removable) {
1087             monitor_printf(mon, " locked=%d", bs->locked);
1088         }
1089         if (bs->drv) {
1090             monitor_printf(mon, " file=");
1091             monitor_print_filename(mon, bs->filename);
1092             if (bs->backing_file[0] != '\0') {
1093                 monitor_printf(mon, " backing_file=");
1094                 monitor_print_filename(mon, bs->backing_file);
1095             }
1096             monitor_printf(mon, " ro=%d", bs->read_only);
1097             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1098             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1099         } else {
1100             monitor_printf(mon, " [not inserted]");
1101         }
1102         monitor_printf(mon, "\n");
1103     }
1104 }
1105
1106 /* The "info blockstats" command. */
1107 void bdrv_info_stats(Monitor *mon)
1108 {
1109     BlockDriverState *bs;
1110
1111     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1112         monitor_printf(mon, "%s:"
1113                        " rd_bytes=%" PRIu64
1114                        " wr_bytes=%" PRIu64
1115                        " rd_operations=%" PRIu64
1116                        " wr_operations=%" PRIu64
1117                        "\n",
1118                        bs->device_name,
1119                        bs->rd_bytes, bs->wr_bytes,
1120                        bs->rd_ops, bs->wr_ops);
1121     }
1122 }
1123
1124 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1125 {
1126     if (bs->backing_hd && bs->backing_hd->encrypted)
1127         return bs->backing_file;
1128     else if (bs->encrypted)
1129         return bs->filename;
1130     else
1131         return NULL;
1132 }
1133
1134 void bdrv_get_backing_filename(BlockDriverState *bs,
1135                                char *filename, int filename_size)
1136 {
1137     if (!bs->backing_hd) {
1138         pstrcpy(filename, filename_size, "");
1139     } else {
1140         pstrcpy(filename, filename_size, bs->backing_file);
1141     }
1142 }
1143
1144 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1145                           const uint8_t *buf, int nb_sectors)
1146 {
1147     BlockDriver *drv = bs->drv;
1148     if (!drv)
1149         return -ENOMEDIUM;
1150     if (!drv->bdrv_write_compressed)
1151         return -ENOTSUP;
1152     if (bdrv_check_request(bs, sector_num, nb_sectors))
1153         return -EIO;
1154     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1155 }
1156
1157 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1158 {
1159     BlockDriver *drv = bs->drv;
1160     if (!drv)
1161         return -ENOMEDIUM;
1162     if (!drv->bdrv_get_info)
1163         return -ENOTSUP;
1164     memset(bdi, 0, sizeof(*bdi));
1165     return drv->bdrv_get_info(bs, bdi);
1166 }
1167
1168 int bdrv_put_buffer(BlockDriverState *bs, const uint8_t *buf, int64_t pos, int size)
1169 {
1170     BlockDriver *drv = bs->drv;
1171     if (!drv)
1172         return -ENOMEDIUM;
1173     if (!drv->bdrv_put_buffer)
1174         return -ENOTSUP;
1175     return drv->bdrv_put_buffer(bs, buf, pos, size);
1176 }
1177
1178 int bdrv_get_buffer(BlockDriverState *bs, uint8_t *buf, int64_t pos, int size)
1179 {
1180     BlockDriver *drv = bs->drv;
1181     if (!drv)
1182         return -ENOMEDIUM;
1183     if (!drv->bdrv_get_buffer)
1184         return -ENOTSUP;
1185     return drv->bdrv_get_buffer(bs, buf, pos, size);
1186 }
1187
1188 /**************************************************************/
1189 /* handling of snapshots */
1190
1191 int bdrv_snapshot_create(BlockDriverState *bs,
1192                          QEMUSnapshotInfo *sn_info)
1193 {
1194     BlockDriver *drv = bs->drv;
1195     if (!drv)
1196         return -ENOMEDIUM;
1197     if (!drv->bdrv_snapshot_create)
1198         return -ENOTSUP;
1199     return drv->bdrv_snapshot_create(bs, sn_info);
1200 }
1201
1202 int bdrv_snapshot_goto(BlockDriverState *bs,
1203                        const char *snapshot_id)
1204 {
1205     BlockDriver *drv = bs->drv;
1206     if (!drv)
1207         return -ENOMEDIUM;
1208     if (!drv->bdrv_snapshot_goto)
1209         return -ENOTSUP;
1210     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1211 }
1212
1213 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1214 {
1215     BlockDriver *drv = bs->drv;
1216     if (!drv)
1217         return -ENOMEDIUM;
1218     if (!drv->bdrv_snapshot_delete)
1219         return -ENOTSUP;
1220     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1221 }
1222
1223 int bdrv_snapshot_list(BlockDriverState *bs,
1224                        QEMUSnapshotInfo **psn_info)
1225 {
1226     BlockDriver *drv = bs->drv;
1227     if (!drv)
1228         return -ENOMEDIUM;
1229     if (!drv->bdrv_snapshot_list)
1230         return -ENOTSUP;
1231     return drv->bdrv_snapshot_list(bs, psn_info);
1232 }
1233
1234 #define NB_SUFFIXES 4
1235
1236 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1237 {
1238     static const char suffixes[NB_SUFFIXES] = "KMGT";
1239     int64_t base;
1240     int i;
1241
1242     if (size <= 999) {
1243         snprintf(buf, buf_size, "%" PRId64, size);
1244     } else {
1245         base = 1024;
1246         for(i = 0; i < NB_SUFFIXES; i++) {
1247             if (size < (10 * base)) {
1248                 snprintf(buf, buf_size, "%0.1f%c",
1249                          (double)size / base,
1250                          suffixes[i]);
1251                 break;
1252             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1253                 snprintf(buf, buf_size, "%" PRId64 "%c",
1254                          ((size + (base >> 1)) / base),
1255                          suffixes[i]);
1256                 break;
1257             }
1258             base = base * 1024;
1259         }
1260     }
1261     return buf;
1262 }
1263
1264 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1265 {
1266     char buf1[128], date_buf[128], clock_buf[128];
1267 #ifdef _WIN32
1268     struct tm *ptm;
1269 #else
1270     struct tm tm;
1271 #endif
1272     time_t ti;
1273     int64_t secs;
1274
1275     if (!sn) {
1276         snprintf(buf, buf_size,
1277                  "%-10s%-20s%7s%20s%15s",
1278                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1279     } else {
1280         ti = sn->date_sec;
1281 #ifdef _WIN32
1282         ptm = localtime(&ti);
1283         strftime(date_buf, sizeof(date_buf),
1284                  "%Y-%m-%d %H:%M:%S", ptm);
1285 #else
1286         localtime_r(&ti, &tm);
1287         strftime(date_buf, sizeof(date_buf),
1288                  "%Y-%m-%d %H:%M:%S", &tm);
1289 #endif
1290         secs = sn->vm_clock_nsec / 1000000000;
1291         snprintf(clock_buf, sizeof(clock_buf),
1292                  "%02d:%02d:%02d.%03d",
1293                  (int)(secs / 3600),
1294                  (int)((secs / 60) % 60),
1295                  (int)(secs % 60),
1296                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1297         snprintf(buf, buf_size,
1298                  "%-10s%-20s%7s%20s%15s",
1299                  sn->id_str, sn->name,
1300                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1301                  date_buf,
1302                  clock_buf);
1303     }
1304     return buf;
1305 }
1306
1307
1308 /**************************************************************/
1309 /* async I/Os */
1310
1311 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1312                                  QEMUIOVector *qiov, int nb_sectors,
1313                                  BlockDriverCompletionFunc *cb, void *opaque)
1314 {
1315     BlockDriver *drv = bs->drv;
1316     BlockDriverAIOCB *ret;
1317
1318     if (!drv)
1319         return NULL;
1320     if (bdrv_check_request(bs, sector_num, nb_sectors))
1321         return NULL;
1322
1323     ret = drv->bdrv_aio_readv(bs, sector_num, qiov, nb_sectors,
1324                               cb, opaque);
1325
1326     if (ret) {
1327         /* Update stats even though technically transfer has not happened. */
1328         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1329         bs->rd_ops ++;
1330     }
1331
1332     return ret;
1333 }
1334
1335 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1336                                   QEMUIOVector *qiov, int nb_sectors,
1337                                   BlockDriverCompletionFunc *cb, void *opaque)
1338 {
1339     BlockDriver *drv = bs->drv;
1340     BlockDriverAIOCB *ret;
1341
1342     if (!drv)
1343         return NULL;
1344     if (bs->read_only)
1345         return NULL;
1346     if (bdrv_check_request(bs, sector_num, nb_sectors))
1347         return NULL;
1348
1349     ret = drv->bdrv_aio_writev(bs, sector_num, qiov, nb_sectors,
1350                                cb, opaque);
1351
1352     if (ret) {
1353         /* Update stats even though technically transfer has not happened. */
1354         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1355         bs->wr_ops ++;
1356     }
1357
1358     return ret;
1359 }
1360
1361 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1362 {
1363     acb->pool->cancel(acb);
1364 }
1365
1366
1367 /**************************************************************/
1368 /* async block device emulation */
1369
1370 typedef struct BlockDriverAIOCBSync {
1371     BlockDriverAIOCB common;
1372     QEMUBH *bh;
1373     int ret;
1374     /* vector translation state */
1375     QEMUIOVector *qiov;
1376     uint8_t *bounce;
1377     int is_write;
1378 } BlockDriverAIOCBSync;
1379
1380 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1381 {
1382     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1383     qemu_bh_delete(acb->bh);
1384     qemu_aio_release(acb);
1385 }
1386
1387 static AIOPool bdrv_em_aio_pool = {
1388     .aiocb_size         = sizeof(BlockDriverAIOCBSync),
1389     .cancel             = bdrv_aio_cancel_em,
1390 };
1391
1392 static void bdrv_aio_bh_cb(void *opaque)
1393 {
1394     BlockDriverAIOCBSync *acb = opaque;
1395
1396     if (!acb->is_write)
1397         qemu_iovec_from_buffer(acb->qiov, acb->bounce, acb->qiov->size);
1398     qemu_vfree(acb->bounce);
1399     acb->common.cb(acb->common.opaque, acb->ret);
1400     qemu_bh_delete(acb->bh);
1401     qemu_aio_release(acb);
1402 }
1403
1404 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1405                                             int64_t sector_num,
1406                                             QEMUIOVector *qiov,
1407                                             int nb_sectors,
1408                                             BlockDriverCompletionFunc *cb,
1409                                             void *opaque,
1410                                             int is_write)
1411
1412 {
1413     BlockDriverAIOCBSync *acb;
1414
1415     acb = qemu_aio_get(&bdrv_em_aio_pool, bs, cb, opaque);
1416     acb->is_write = is_write;
1417     acb->qiov = qiov;
1418     acb->bounce = qemu_blockalign(bs, qiov->size);
1419
1420     if (!acb->bh)
1421         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1422
1423     if (is_write) {
1424         qemu_iovec_to_buffer(acb->qiov, acb->bounce);
1425         acb->ret = bdrv_write(bs, sector_num, acb->bounce, nb_sectors);
1426     } else {
1427         acb->ret = bdrv_read(bs, sector_num, acb->bounce, nb_sectors);
1428     }
1429
1430     qemu_bh_schedule(acb->bh);
1431
1432     return &acb->common;
1433 }
1434
1435 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
1436         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1437         BlockDriverCompletionFunc *cb, void *opaque)
1438 {
1439     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1440 }
1441
1442 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
1443         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1444         BlockDriverCompletionFunc *cb, void *opaque)
1445 {
1446     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 1);
1447 }
1448
1449 /**************************************************************/
1450 /* sync block device emulation */
1451
1452 static void bdrv_rw_em_cb(void *opaque, int ret)
1453 {
1454     *(int *)opaque = ret;
1455 }
1456
1457 #define NOT_DONE 0x7fffffff
1458
1459 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1460                         uint8_t *buf, int nb_sectors)
1461 {
1462     int async_ret;
1463     BlockDriverAIOCB *acb;
1464     struct iovec iov;
1465     QEMUIOVector qiov;
1466
1467     async_ret = NOT_DONE;
1468     iov.iov_base = (void *)buf;
1469     iov.iov_len = nb_sectors * 512;
1470     qemu_iovec_init_external(&qiov, &iov, 1);
1471     acb = bdrv_aio_readv(bs, sector_num, &qiov, nb_sectors,
1472         bdrv_rw_em_cb, &async_ret);
1473     if (acb == NULL)
1474         return -1;
1475
1476     while (async_ret == NOT_DONE) {
1477         qemu_aio_wait();
1478     }
1479
1480     return async_ret;
1481 }
1482
1483 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1484                          const uint8_t *buf, int nb_sectors)
1485 {
1486     int async_ret;
1487     BlockDriverAIOCB *acb;
1488     struct iovec iov;
1489     QEMUIOVector qiov;
1490
1491     async_ret = NOT_DONE;
1492     iov.iov_base = (void *)buf;
1493     iov.iov_len = nb_sectors * 512;
1494     qemu_iovec_init_external(&qiov, &iov, 1);
1495     acb = bdrv_aio_writev(bs, sector_num, &qiov, nb_sectors,
1496         bdrv_rw_em_cb, &async_ret);
1497     if (acb == NULL)
1498         return -1;
1499     while (async_ret == NOT_DONE) {
1500         qemu_aio_wait();
1501     }
1502     return async_ret;
1503 }
1504
1505 void bdrv_init(void)
1506 {
1507     module_call_init(MODULE_INIT_BLOCK);
1508 }
1509
1510 void *qemu_aio_get(AIOPool *pool, BlockDriverState *bs,
1511                    BlockDriverCompletionFunc *cb, void *opaque)
1512 {
1513     BlockDriverAIOCB *acb;
1514
1515     if (pool->free_aiocb) {
1516         acb = pool->free_aiocb;
1517         pool->free_aiocb = acb->next;
1518     } else {
1519         acb = qemu_mallocz(pool->aiocb_size);
1520         acb->pool = pool;
1521     }
1522     acb->bs = bs;
1523     acb->cb = cb;
1524     acb->opaque = opaque;
1525     return acb;
1526 }
1527
1528 void qemu_aio_release(void *p)
1529 {
1530     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1531     AIOPool *pool = acb->pool;
1532     acb->next = pool->free_aiocb;
1533     pool->free_aiocb = acb;
1534 }
1535
1536 /**************************************************************/
1537 /* removable device support */
1538
1539 /**
1540  * Return TRUE if the media is present
1541  */
1542 int bdrv_is_inserted(BlockDriverState *bs)
1543 {
1544     BlockDriver *drv = bs->drv;
1545     int ret;
1546     if (!drv)
1547         return 0;
1548     if (!drv->bdrv_is_inserted)
1549         return 1;
1550     ret = drv->bdrv_is_inserted(bs);
1551     return ret;
1552 }
1553
1554 /**
1555  * Return TRUE if the media changed since the last call to this
1556  * function. It is currently only used for floppy disks
1557  */
1558 int bdrv_media_changed(BlockDriverState *bs)
1559 {
1560     BlockDriver *drv = bs->drv;
1561     int ret;
1562
1563     if (!drv || !drv->bdrv_media_changed)
1564         ret = -ENOTSUP;
1565     else
1566         ret = drv->bdrv_media_changed(bs);
1567     if (ret == -ENOTSUP)
1568         ret = bs->media_changed;
1569     bs->media_changed = 0;
1570     return ret;
1571 }
1572
1573 /**
1574  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1575  */
1576 int bdrv_eject(BlockDriverState *bs, int eject_flag)
1577 {
1578     BlockDriver *drv = bs->drv;
1579     int ret;
1580
1581     if (bs->locked) {
1582         return -EBUSY;
1583     }
1584
1585     if (!drv || !drv->bdrv_eject) {
1586         ret = -ENOTSUP;
1587     } else {
1588         ret = drv->bdrv_eject(bs, eject_flag);
1589     }
1590     if (ret == -ENOTSUP) {
1591         if (eject_flag)
1592             bdrv_close(bs);
1593         ret = 0;
1594     }
1595
1596     return ret;
1597 }
1598
1599 int bdrv_is_locked(BlockDriverState *bs)
1600 {
1601     return bs->locked;
1602 }
1603
1604 /**
1605  * Lock or unlock the media (if it is locked, the user won't be able
1606  * to eject it manually).
1607  */
1608 void bdrv_set_locked(BlockDriverState *bs, int locked)
1609 {
1610     BlockDriver *drv = bs->drv;
1611
1612     bs->locked = locked;
1613     if (drv && drv->bdrv_set_locked) {
1614         drv->bdrv_set_locked(bs, locked);
1615     }
1616 }
1617
1618 /* needed for generic scsi interface */
1619
1620 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1621 {
1622     BlockDriver *drv = bs->drv;
1623
1624     if (drv && drv->bdrv_ioctl)
1625         return drv->bdrv_ioctl(bs, req, buf);
1626     return -ENOTSUP;
1627 }
1628
1629 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1630         unsigned long int req, void *buf,
1631         BlockDriverCompletionFunc *cb, void *opaque)
1632 {
1633     BlockDriver *drv = bs->drv;
1634
1635     if (drv && drv->bdrv_aio_ioctl)
1636         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1637     return NULL;
1638 }
1639
1640 void *qemu_blockalign(BlockDriverState *bs, size_t size)
1641 {
1642     return qemu_memalign((bs && bs->buffer_alignment) ? bs->buffer_alignment : 512, size);
1643 }